MASK-GUIDED COARSE-TO-REFINED PATCH-WISE 3D CNN FOR CORONARY ARTERY SEGMENTATION

Abstract

Accurate segmentation of coronary arteries from 3D computed tomography angiography (CTA) remains a challenging problem. This difficulty arises due to high volumetric resolution, severe class imbalance, and the thin, elongated structure of coronary vessels. Conventional full-volume 3D convolutional neural networks (3D-CNNs) are computationally expensive and often fail to preserve fine vessel details. To address these challenges, this paper proposes a Mask-Guided Coarse-to-Refined Patch-Wise 3D CNN for coronary artery segmentation. High-resolution CTA volumes are first decomposed into fixed-size 3D patches to reduce memory consumption while preserving local anatomical information. A lightweight coarse segmentation network is then employed to generate an initial probabilistic localization of coronary arteries. Unlike existing approaches that use coarse predictions only as auxiliary outputs, the proposed method introduces a mask-guided feature modulation mechanism. This mechanism uses coarse vessel probability maps as structural priors to enhance vessel-relevant features during refinement. In addition, vessel-aware attention gates are integrated to suppress background noise and emphasize anatomically salient vascular regions. The refined segmentation network adopts residual learning and attention-guided decoding to produce voxel-wise artery probability maps. To handle class imbalance and preserve vessel continuity, a hybrid loss function combining binary cross-entropy, Dice loss, and a topology-aware constraint is employed. Experimental evaluation demonstrates the effectiveness of the proposed approach. Using patch-wise inputs of size 128 × 128 × 128, the proposed method achieves a Dice score of 94.10% for coronary artery segmentation. These results confirm that explicit mask-guided refinement significantly improves segmentation accuracy and vessel continuity.